Modeling the initial state of early phonetic learning in infants

Installation

Clone the repository and install the package with pip:

git clone https://github.com/mxmpl/initial-phonetic-learning
cd initial-phonetic-learning
pip install .

If you want to train models, evaluate them or recreate the features for the t-SNE, you must install the CPC3 package, with pytorch and torchaudio.

Usage

This repository provides a CLI to build jobs to train and evaluate models with CPC3 in the context of this project, and utilities to analyze the results and reproduce the figures of the paper.

Downloads

Training data:

• Ambient sounds: 54G, md5sum: a625f3564a7697cf827d478118045c1c
• English training set: 53G, md5sum: 8de965e97a6dbc73fcfdc994fceed750

Results:

• ABX scores: 19M, md5sum: 4afecc9416484af4e7fd61483012ee6c
• t-SNE data: 401M, md5sum: 13193518acb207f00858177eefc15ffb